《人工智能（智能系统指南，第二版）》读书笔记——9、第八章

1. introduction to knowledge-based intelligent systems(summary / questions for review / references)

2. rule-based expert systems

3. uncertainty management in rule-based expert systems

4. fuzzy expert systems

5. frame-based expert systems

6. artificial neural networks

7. evolutionary computation

8. hybrid intelligent systems

9. knowledge engineering and data mining

8. hybrid intelligent systems

Hybird intelligent systems are systems that combine at least two intelligent technologies; for example, a combination of a neural network and a fuzzy system results in a hybrid neuro-fuzzy system.

Probabilistic reasoning, fuzzy set theory, neural networks and evolutionary computation form the core of soft computing(SC, deals with soft values, or fuzzy sets; is capable of operating with uncertain, imprecise and incomplete information in a manner that reflects human thinking), an emerging approach to building hybrid intelligent systems capable of reasoning and learning in uncertain and imprecise environments.

Probabilistic reasoning is mainly concerned with uncertainty, fuzzy logic with imprecision, neural networks with learning, and evolutionary computation with optimisation. Table 8.1 presents a comparison of different intelligent technologies.

Both expert systems and neural networks attempts to emulate human intelligence, but use different means. While expert systems rely on IF-THEN rules and logical inference, neural networks use parallel data processing. An expert system cannot learn, but can explain its reasoning, while a neural network can learn, but acts as a black-box. These qualities make them good candidates for building a hybrid intelligent system, called a neural or connectionist expert system.

Neural expert systems use a trained neural network in place of the knowledge base. Unlike conventional rule-based systems, neural expert systems can deal with noisy and incomplete data(approximate reasoning). Domain knowledge can be utilised in an initial struture of the neural knowledge base. After training, the neural knowledge base can be interpreted as a set of IF-THEN production rules.

A neuro-fuzzy system corresponding to the Mamdani fuzzy inference model can be represented by a feedforward neural network consisting of five layers: input, fuzzification, fuzzy rule, output membership and defuzzification.

An adaptive neuro-fuzzy inference system, ANFIS(Roger Jang, Tsing Hua University), corresponds to the first-order Sugeno fuzzy model. The ANFIS is represented by a neural network with six layers: input, fuzzification, fuzzy rule, normalisation, defuzzification and summation.

pic

The ANFIS uses a hybrid learning algorithm that combines the least-squares estimator(最小乘方估算) with the gradient descent method(梯度下降法). In the forword pass, a training set of input patterns is presented, neuron outputs are calculated on a layer-by-layer basis(输出一层接一层的计算), and rule consequent parameters are identified by the least-squares estimator. In the backward pass, the error signals are propagated back and the rule antecedent parameters are updated according to the chain rule(链接规则).

Genetic algorithms are effective for optimising weights and selecting the topology of a neural network.

Evolutinary computation can also be used for selecting an appropriate set of fuzzy rules for solving a complex classification problem. While a complete set of fuzzy rules for solving a complex classification problem. While a complete set of fuzzy IF-THEN rules is generated from numerical data by suing multiple fuzzy rule tables, a genetic algorithm is used to select a relatively small number of fuzzy rules with high classification power.